Abstract
The vibration signals measured in practical engineering are usually complex and noisy, which brings challenges to fault diagnosis. In addition, industrial scenarios also put forward higher requirements for the accuracy and computational efficiency of diagnostic models. Aiming at these problems, an improved multiscale branching convolutional neural network is proposed for rolling bearing fault diagnosis. The proposed method first applies the multiscale feature learning strategy to extract rich and compelling fault information from diverse and complex vibration signals. Further, the lightweight dynamic separable convolution is elaborated and coupled into the feature extractor to "slim down" the model, reduce the computational loss on the one hand, and further improve the model's adaptive learning ability for different inputs hand. Extensive experiments indicate that the proposed method is significantly improved compared with existing multi-scale neural networks.